Handbook

Genesys Cloud quality management: evaluations, calibration and AI scoring

A detailed handbook for the people who run and are measured by Genesys Cloud quality management. It covers forms and their limits, critical and fatal questions, weighting and N/A, versions and divisions, sampling policies, scoring, release, disputes, calibration, Evaluation Assistance, AI scoring, auto-evaluation, reporting, coaching, retention, permissions and licensing. Checked against Genesys sources on 16 November 2026.

By Dave Tidwell, Managing Director36 min read

In short

  • A score is weighted by group, ignores N/A questions entirely, and a fatal No zeroes both the overall and critical score.
  • Published forms cannot be edited and their division cannot change, so settle divisions and versioning before publishing.
  • By-agents policies evaluate the last agent and the other policy types the first, so transferred calls are sampled differently.
  • AI scoring reads the form text and the transcript only, hides answers below 75% historical accuracy and changes when models do.
  • Evaluations are kept indefinitely but recordings are not, so align retention before a score has to be defended.

01

What quality management is for, and who does what

This handbook is for the people who design, run and are measured by quality management in Genesys Cloud: quality managers, evaluators, supervisors, workforce and HR leads, and the administrators who configure it. Every statement about Genesys Cloud comes from Genesys's documentation, listed in the sources and checked on 16 November 2026. Where something is our own practice, it is labelled QVCCS practice. Our workforce engagement management handbook covers quality as one part of the suite; this handbook goes into the detail.

1.1 Three outputs, and the one that gets forgotten

Genesys describes quality management as a framework and toolset for making sure teams deliver the expected quality and empathy. By evaluating and scoring conversations, a quality team can identify agents who deserve recognition, agents who need training or coaching, and situations where a process or script should change.

Read that as three outputs: recognition, development, and a change to the process or the script. The third is the one that pays for the programme. When evaluations keep finding agents struggling with the same step, that is usually evidence about the step, not about the agents.

Genesys groups quality with four neighbouring capabilities: quality administration (forms, policies and security), evaluations (scoring, calibration and disputes), coaching, and speech and text analytics, which Genesys says analyses every interaction. That last neighbour is why this handbook is long. Analytics is now how evaluations are pre-filled, and a quality programme that does not understand analytics is configuring blind.

1.2 Why quality needs more care than most features

Quality is the Genesys Cloud capability whose output most often becomes evidence: in a disciplinary process, an employment tribunal, a regulated complaint or an auditor's file. Part 13 follows from that. Genesys keeps evaluation data indefinitely, but the recordings that justify a score follow a retention policy you control, and the two clocks do not match.

1.3 The roles

Quality management roles (from Genesys)
RoleWhat Genesys says it does
Quality administratorManages policies, creates and publishes evaluation forms, assigns evaluations, coordinates calibration and can view all quality activity
Quality evaluatorReviews recordings, scores evaluations and takes part in calibration
Expert evaluatorBecomes the benchmark for scoring an interaction in a calibration
CalibratorCollects and compares calibration scores; typically the quality administrator, but Genesys says an agent or evaluator can be specified
AgentReviews released evaluations, adds comments where possible, acknowledges review and can dispute
  • The expert evaluator is a designation made per calibration rather than a permission, and it is the most consequential role in the feature: the expert's own leanings become the team's standard (part 10).
  • Release, review and dispute each carry their own permission, which is why they appear as separate steps throughout this handbook.

02

The workflow, end to end

Genesys documents the whole sequence on its quality assurance and compliance use-case page. Two steps in it are structural rather than incidental, and recording policy sits at the top of it.

2.1 The documented sequence

  1. An administrator assigns roles; the standard users are Quality Admin and Evaluator.
  2. The quality admin creates evaluation forms (name, question groups, weighting, group properties, questions and their types) and publishes a form to make it available for assigning evaluations.
  3. A supervisor can also create an evaluation by hand against a single interaction.
  4. For each question, automation is set to No Automation, AI scoring or Evaluation Assistance.
  5. The quality admin adds evaluation actions to policies for agent groups or individuals, choosing interaction type, criteria, forms, evaluators, a count per period, and whether the evaluation is for the agent or for calibration.
  6. The evaluator is notified, and the evaluator dashboard shows assigned and completed work.
  7. The evaluator releases the evaluation for the agent to review, and the agent is notified.
  8. The agent reviews it under Performance, My Performance, Evaluations, acknowledges it, can comment and can ask for further review, after which an evaluator can rescore.
  9. Evaluators and supervisors see completed evaluations under Performance, Agents, Evaluations.

2.2 Two structural points

Release is a deliberate step, not a consequence of scoring. It has its own action and its own permission. An evaluation can be complete and invisible to the agent, which is sometimes exactly right and sometimes a quiet failure of the programme (part 9).

Calibration shares the sampling machinery. Because a policy action can assign for agent evaluation or for calibration, a sampling design that forgets calibration will never produce the consistency data that makes scores defensible (part 10).

2.3 Recording policy comes first

Genesys's policy actions include deleting, retaining, archiving or exporting a recording, starting screen recording, assigning for evaluation or calibration and starting a survey. Genesys states that if the policy action is to delete the recording, no other actions can be applied to the interaction. An interaction deleted by policy cannot be evaluated, which is the commonest reason an evaluation policy appears to have stopped working.

03

Evaluation forms: anatomy, limits and question types

A form is three levels deep, and the boundaries between levels are where all the behaviour lives. Genesys publishes the limits in the form creation article itself, which tells you they are reached in practice.

3.1 Form, group, question, answer

The structure of a Genesys Cloud evaluation form
LevelWhat it carries
FormName, type, division, dispute and revision settings, and scaled scoring
Question groupName, weight or scaling factor, N/A enabled, Answers Default To and Display Conditionally
QuestionText, type, whole-number values, N/A, critical, fatal, comments required, visibility conditions, help text, automation mode and evaluation focus
AnswerThe options an evaluator, or automation, can select
  • Genesys describes a group as a label that categorises a set of questions, such as Product Knowledge. Groups are also the unit of weighting and of group-level N/A, so the group structure is the most important design decision on a form.
  • Genesys says forms measure script adherence, compliance with business practices, customer satisfaction or other benchmarks. Those need different question styles and carry different confidence, which is why we design one form per purpose.
  • A form is created as one of two types: Evaluation Form or Agent Auto-Evaluation (part 12). Decide the type before writing any questions.

3.2 The published limits

Evaluation form limits (from Genesys, checked 16 November 2026)
LimitValueWhere Genesys states it
Form name250 charactersCreate and publish an evaluation form
Questions per form200Create and publish an evaluation form
Question group name250 charactersCreate and publish an evaluation form
Question text500 charactersCreate and publish an evaluation form; automated scoring options
Help text, on AI-scored questions6,000 charactersAutomated scoring options
Each answer, on AI-scored questions500 charactersAutomated scoring options
Topics in one Evaluation Assistance condition10Create and publish an evaluation form
AI-scored questions per form50Create and publish an evaluation form
Answer options on an AI-scored question5Automated scoring options

3.3 Designing around the limits

200 questions is generous, and that is the trap. The limit protects nothing; the arithmetic of a quality team's week does. Genesys's own sample form for AI scoring has ten questions: greeting, identity verification, active listening, personalisation, empathy, resolution, escalation process, compliance, hold management and closing.

50 AI-scored questions is a limit on how much of a form can be automated, not on its size. A 60-question form can never be fully AI-scored. And 6,000 characters of help text is where the real scoring guidance goes, because question text stops at 500. Genesys publishes no limit on the number of forms in an organisation, so confirm it before a design depends on one.

3.4 Group properties and question types

Each group has three documented properties. N/A Enabled lets the evaluator skip every question in the group. Answers Default To starts each answer at the highest scoring, lowest scoring, a user-defined option or N/A; Genesys adds that for multi-select questions, highest or lowest selects all applicable options, and that where both a scored default and N/A are possible, the default becomes N/A. Display Conditionally sets criteria for whether the group is shown.

Genesys lists multi-select, multiple choice, range and yes/no questions, plus custom types where available. Question values are whole numbers, so half-point weighting has to be modelled as 5 and 10. One asymmetry matters before designing an automated form: for multi-select questions, only Evaluation Assistance is supported, never AI scoring.

The editor helps. A table of contents builds as you add groups; a group or question that is not properly configured gets a red border and Save stays unavailable; and a preview lets you check the form before publishing. Genesys also notes that to move a question group, all of its answers must be complete.

04

Critical and fatal questions

Two question properties look similar and behave very differently, and both change what a score means. Genesys's worked examples show the arithmetic.

4.1 Critical questions and the critical score

Genesys defines a critical question as one that is critical to the success of an interaction. Critical questions must be scored Yes or No, and they count towards both the overall score and a separate critical score, which reflects the agent's compliance with the critical questions. The critical score sums each critical question's points divided by the total critical points available.

In Genesys's critical example, a form has two groups of two one-point questions, weighted 40 and 10, with two questions marked critical. An agent who scores one of two in each group, and one of the two critical questions, receives 50% overall and a critical score of 50%.

4.2 Fatal questions

Genesys defines a fatal question as one which, if scored No, results in an evaluation score of zero, and its glossary adds that fatal questions are also critical questions. In the fatal example, the same form marks question 2.2 as fatal and it is answered No. Genesys states the result as 0% overall and 0% critical.

An agent who answered three of four questions correctly scores zero on both measures. That is the intent for a real compliance failure, such as a required disclosure not given or identity not verified, and it is also the fastest way to make a quality programme resented if fatal is applied to anything that merely matters.

One editorial point. Genesys's printed formula for the fatal example does not compute as written; read as bracketed it multiplies part of the denominator by zero. The outcome, zero, is stated and matches the definition, so rely on the rule and the result rather than reproducing the expression.

4.3 Document the consequence

A fatal answer collapses an agent's monthly average, and possibly an incentive. That is a policy decision about people, not a form setting.

05

Weights, N/A and defaults

Three settings decide what a percentage actually means: group weights, N/A, and default answers. Each is documented, and each produces a score that surprises people who have not read the documentation.

5.1 The weighting formula

Weights belong to question groups. Genesys says they multiply the total maximum points of the questions in each group, and the score is the sum of each group's points times its weight, divided by the sum of each group's possible points times its weight.

Genesys's plain example: group 1 has two one-point questions weighted 40, and group 2 has one one-point question weighted 20. The agent scores one of two in group 1 and one of one in group 2, and the result is 60%. The agent answered two of three questions correctly, 67%, and scored 60%, because the group they lost a point in carries twice the weight. That is correct behaviour, and it is the first thing to explain to an agent reading a score: the percentage is not the share of questions answered correctly.

Genesys's use-case page says group weighting defaults to even and always rebalances to the full total when adjusted. Because the formula works on ratios, weights of 40 and 20 behave like 2 and 1.

5.2 Scaled scoring, and weights that come back

Forms can use scaled scoring, with a Group Score Scaling Factor per group. Genesys's edit article documents a behaviour behind a real support ticket: if you turn scaled scoring off and then on again, previous weight values are restored, and if none exist, all weights are set to 1. Switching scaled scoring is not a round trip. Treat the toggle as a scoring-model change, check the weights afterwards, and never flip it as an experiment on a live form.

5.3 N/A leaves the arithmetic

Genesys states that questions or groups marked N/A are not included in the score, and show zero out of zero possible points. When a whole group is marked N/A, Genesys adjusts the remaining groups: each group's weight is divided by the sum of the non-N/A group weights. In Genesys's example, two groups weighted 25 each, with the first group N/A, leave the second group carrying all the weight, and one of two correct gives 50%.

So the score is computed over whatever was scored. Two agents can both receive 90% on the same form for very different amounts of assessed behaviour, and a monthly average can mix percentages taken over 14 questions with percentages taken over 9.

  • N/A is an evaluator decision, not a rule, so calibration measures N/A discipline as well as scoring (part 10).
  • Group-level N/A is a large lever: one click and a weighted slice of the form disappears.
  • A group that defaults to the highest score with N/A enabled can default to N/A, silently shrinking the form whenever the evaluator does not look.
  • A critical question marked N/A contributes nothing to the critical score, so an agent who was not assessed on a mandatory step can still carry a perfect critical score. Genesys documents no flag for a critical question left N/A.

5.4 Defaults and conditional display

Answers Default To exists to save time, and every default is a policy position. Default to highest says the agent passes unless the evaluator intervenes: fast, and biased towards yes, which is wrong for compliance questions. Default to lowest is slower and biased towards findings. On Agent Auto-Evaluation forms the default becomes the fallback when automation cannot answer, so there it records uncertainty as a result (part 12).

Conditional display works at both levels: Display Conditionally on groups, and visibility conditions on questions. Used well, it keeps a form short. Used carelessly, it means two evaluators see different forms for the same kind of interaction and score them incomparably.

06

Publishing, versioning and divisions

Genesys models forms as published versions that cannot be edited, and makes division a permanent choice at publication. Both rules are tidy, and both punish late decisions.

6.1 Published forms are read-only

Genesys stores saved and published versions of a form. You can edit any saved form, but not a published one; you create a new version based on the last published version. A published form can be reviewed in read-only mode, where you interact with it as an evaluator would without saving, which makes it a safe place for evaluator training.

The propagation rule is the one to plan around: if policies use the form, new interactions use the new published version. Each form also carries its last revision date as a version identifier, and Genesys added version ID tracking on the interaction's details page in June 2022. So an evaluation permanently records which version of the form produced it, and a March score can be read against March criteria, which is what an appeal needs.

6.2 Copying is branching

A copy is saved as Copy of the original name, holds everything from the original (groups, questions, conditionality and weighting), does not keep version information, starts as version one and stays unpublished until you publish it. Unpublished forms can be copied too, and a copy whose name would breach the 250-character limit prompts for a new title.

  • Never edit a form mid-period to fix a typo without announcing it. The typo is cosmetic; an unannounced change of criteria is a fairness problem.
  • Give every new version a communication date, so agents know the criteria changed and from when.
  • Rename copies deliberately, because the name appears in every view, filter and export.
  • Check which policies use a form before publishing a new version, and plan versions around reporting periods: a new version breaks question-level trend lines.

6.3 One form, one division, for good

Genesys states that a form can belong to only one division, that you can choose only divisions you have permission for, and that forms created before division support are assigned to Unassigned. After a form is published, its division cannot change; to use a different division you create a new version before publishing.

A group-structured estate, with a division per brand, market or outsourcer, therefore needs a form per division or one form in a parent division. Copies make per-division forms easy to create and hard to keep in step. Unassigned forms are a migration trap on older estates: depending on how roles are scoped, a division's own quality team may not see them, and the fix is recreation rather than a change. Settle the division architecture, as our roles, permissions and divisions handbook explains, before a single form is drafted.

07

Sampling: policies, and which agent is evaluated

Sampling decides whether a quality programme is statistically meaningful or anecdotal. Genesys offers four policy actions with different meanings, and a rule about which agent is evaluated that changes between them.

7.1 Four policy actions, three questions

Evaluation policy actions (from Genesys)
ActionWhat it doesGenesys's use caseThe question it answers
Assign Evaluations by InteractionsAn evaluation for every matching interaction; several evaluators each get the same setNew hires, or compliance monitoring where every interaction with potential issues must be evaluatedHas every interaction of this type been assessed?
Assign Evaluations by EvaluatorsLimits how many evaluations are assigned, per the matching criteriaA set number per evaluator per interval, for example 10, regardless of how many interactions matchDoes each evaluator have a manageable queue?
Assign Evaluations by AgentsA specified number per period for each agent in the matching criteriaAn even distribution, for example two per agent per weekHas each agent been assessed enough, and fairly?
Assign Calibration EvaluationsFeeds calibration rather than agent assessmentConsistency checks between evaluatorsDo our evaluators agree?

7.2 The by-agents action in detail

Genesys documents more about the by-agents action than any other. The maximums are 50 evaluations per agent per day, 175 per week and 700 per month. Sampling can be a fixed number per period, or a percentage of interactions with a minimum. You can set an agent-connected duration, measured from the agent joining to disconnecting, including hold and excluding after-call work and dialling. Evaluations are spread evenly among the chosen evaluators, with any remainder distributed at random, and created through the period, usually at the start of the next day in the policy's time zone.

  • Periods are calendar units, so a policy created on a Wednesday creates about half the weekly count for that first week.
  • Agents matched through a team or queue rather than named in the policy are prorated from their first matching interaction, and again after long inactivity: two months for monthly policies, five weeks for weekly, 30 days for daily.
  • Adding or removing agents does not reset the policy, but changing other parameters acts like turning it off and on, and previous evaluations no longer count towards the total.
  • No evaluation is created for email or messaging interactions where the customer was the last participant.

7.3 First agent or last agent

Genesys states that for by-evaluators, by-interactions and calibration actions, the policy evaluates the first agent connected to the interaction. For by-agents, it evaluates the last agent who participated and meets the matching criteria, and never a user who was a monitor or coach. Genesys gives no reason for the difference.

In practice, on a call answered by a first-line agent and transferred to a specialist, a by-agents policy evaluates the specialist and a by-interactions policy evaluates the first-line agent. A programme running both, which is most programmes with a compliance trigger alongside routine sampling, evaluates different populations without anyone noticing, and the averages are not comparable. That consequence is our inference from Genesys's rule.

When you create an evaluation by hand, you choose the agent to be evaluated whenever two or more agents took part, so ad-hoc evaluation handles multi-agent interactions by design. Since October 2023, quality managers can also reassign evaluations through scoring, review and release, and agents are notified of assignments regardless of their status.

7.4 The sampling plan is evidence

The first question a tribunal, regulator or union representative asks about a quality score is how the interaction was chosen. Randomly, five per agent per month from a named policy, plus every interaction where a named compliance topic fired, is an answer. The evaluator picked them is not.

08

Creating and scoring evaluations

Policies are the backbone; manual creation is where a programme keeps its judgement. This part covers creating evaluations, the evaluator's tools, and scoring while an interaction is still live.

8.1 Creating an evaluation by hand

Genesys notes that in some organisations evaluators are responsible for choosing interactions. The route is Analytics Workspace, the Interactions view, a filtered interaction, its Quality Summary tab and Create Evaluation, then the form, the agent to be evaluated where several took part, and the evaluator or evaluators. Assign to me assigns it to yourself. For an Agent Auto-Evaluation form you choose between Auto Submit with Virtual Supervisor and requiring human review by assigning evaluators.

Two documented limits apply. You can create at most 50 evaluations for one agent in a 24-hour period, with a message when the limit is reached. And evaluations with AI-scored questions can take up to 20 minutes to generate, which matters when an evaluator opens one straight after creating it.

The filtering is the useful part. Genesys's employee performance use case describes searching interactions by metadata and by speech and text analytics criteria, so an evaluator can find every interaction where a topic fired or sentiment fell, and evaluate on suspicion rather than chance.

8.2 The evaluator's tools

Evaluators work from the quality evaluator dashboard, whose Interactions Needing Attention table lists assigned evaluations not yet completed, and whose Agent Activity table shows how many evaluations they completed and the average score they gave each agent.

Alongside the form sit the recording and transcript, annotations (with their own Recording, Annotation permissions), and comments. Since October 2023 evaluators can add private comments that the agent does not see. For evaluations created from 19 February 2024, comment fields, including private and answer comments, take up to 5,000 characters. An evaluator can schedule release to the agent, but not more than 45 days ahead.

Where AI scoring is enabled, Genesys prefills answers when there is adequate confidence, shows an AI icon by each, and displays the reason the answer was chosen, which can be copied into the comment. Genesys also says plainly that AI scoring sometimes does not align with its own stated reason, and that you can then edit the answer manually.

8.3 Scoring a live interaction

Genesys documents evaluating in-progress interactions through the same Quality Summary route, with one instruction that is easy to miss: to evaluate a call while it is taking place, select your own user account as the evaluator. Genesys's managers' overview confirms that both in-progress and recorded evaluations are supported, alongside real-time monitoring and coaching.

Live evaluation earns its keep in a new agent's first week, on a high-value or sensitive interaction, and for calibrating supervisors' real-time judgement. But a live score is given without rewinding or reflection, so it is harder and less consistent than reviewing a recording.

09

Release, review, disputes and rescoring

This part covers what happens after scoring, where most programmes either earn agents' trust or lose it. It ends with the clocks that run through the whole cycle.

9.1 Release: the gate everyone forgets

Release is its own step, with its own permission, Quality, Evaluation, Release, held by both default quality roles. The Agents Evaluation Summary and Detail views include evaluations whether or not they have been released, which is how you detect a programme stalled at this gate. There are good reasons to hold an evaluation: a manager needs to see a serious finding first, a calibration is in progress, or the interaction is sensitive. The failure is when holding becomes the default, completion numbers look healthy, and no feedback reaches anyone for weeks.

9.2 Agent review and sign-off

Agents open evaluations from My Performance, Evaluations, review the scores and comments with the recording, add comments, and select I have reviewed this evaluation. Genesys states that you cannot review evaluations older than 90 days from the day the evaluation was created, although they are not deleted. The clock runs from creation, not release, so a release backlog eats into the agent's window before they have seen anything.

Agents are notified when an evaluation is released, and the agent activity view's summary card shows recent evaluations, 30 days old or less, that are pending review. After that the prompt goes quiet while the right to review runs on. Whether the agent reviewed and commented is recorded on the interaction's quality summary, and changing sign-off has its own permission, Edit Agent Sign-off.

9.3 Disputes

Disputes arrived in October 2023. Two switches must both be on: Allow disputes on the form's Settings tab, and the Quality, Evaluation, Dispute permission for the person. Genesys states that agents can dispute completed and released evaluations, and that supervisors can dispute completed but unreleased evaluations only if they are the assignee. Agents can open a dispute within 45 days; after that they cannot. A disputed evaluation moves to In Review, and the assignee either rescores it or declines the dispute.

The form also sets the maximum number of disputes per evaluation, which Genesys states is up to 5, and how disputes are assigned: to the original evaluator, to a named individual, or left unassigned for Process Automation to designate the assignee. Genesys's allow-disputes page describes a separate 45-day timeframe for the evaluator to review each dispute.

  • Genesys warns that once disputes are set up, existing API queries return evaluations in In Review and Retracted states, so custom reports and applications may need updating.
  • Once a dispute is resolved, any change in gamification points appears on the agent's scorecard.
  • Assigning disputes to the original evaluator suits misunderstandings about what a question meant, and is a poor route for a contested judgement.

9.4 Rescoring, revisions and retraction

With the form's Allow Revisions setting on and the Edit Score or Edit permission, an evaluator can revise a completed score. Previous versions stay available in an Evaluation versions list, always in Retracted state, with a maximum of 10 revisions; the Revisions column shows the count. Genesys states that when you rescore, the original score disappears and only the new score remains visible, and that scoring results exclude deleted evaluations.

Two edge cases have teeth. If a quality score is a gamification metric and an evaluator rescores after six days, Genesys ignores the rescore for points. And if a rescore is saved but not submitted, the evaluation becomes unreleased and Genesys removes the points given for it, so an abandoned edit looks to the agent like a score change.

9.5 The clocks

Quality management time limits (from Genesys)
ClockLengthMeasured fromWhat happens
Evaluation completion90 daysAssignmentThe evaluation expires, or earlier if the interaction is purged
Agent review90 daysEvaluation creationThe agent can no longer review; the evaluation is kept
Opening a dispute45 daysNot stated in the article textAgents can no longer open a dispute
Evaluator reviewing a dispute45 daysThe disputeDescribed on the allow-disputes page
Scheduled releaseUp to 45 days aheadSubmissionRelease cannot be scheduled further out
Gamification rescore6 daysReleaseLater rescores do not adjust points
  • The quality summary shows a live Time Remaining countdown for the evaluator to complete or delete an evaluation.
  • Evaluations expiring means the sample was cut silently and selectively: the interactions left unevaluated are rarely a random subset.
  • A short interaction retention policy shortens the completion window, and a recording that has gone cannot support an evaluation, a dispute or a disciplinary (part 13).

10

Calibration

Calibration turns individual judgements into a measurement system. Genesys says it standardises evaluations by comparing several reviews of the same interaction, to make scoring consistent and fair, isolate and refine questions with inconsistent scores, set expectations, calibrate against an expert, and measure improvement objectively.

10.1 Running a calibration

From an interaction's Quality Summary, choose Create Calibration, then the form and the evaluators. Under Scoring Index Options, select Expert Evaluator to index scores against an expert's review. The Calibration Results tab summarises the high, low and average scores and the variance between evaluators. Calibrations can also be assigned automatically by policy, using criteria such as queue, call length or agent.

Genesys's Calibration Evaluations dashboard tracks sessions, evaluator activity and scoring trends, and Genesys highlights the Unfinished column for spotting incomplete calibrations. A calibration with missing evaluators produces a variance from an incomplete sample: the number looks valid and is not. Calibration evaluations are excluded from the Agents Evaluation Summary view, which is correct, since they are not an assessment of the agent.

10.2 What calibration really tests

The aim most programmes miss is refining questions. Consistently divergent scores on one question usually mean the question is ambiguous, and the fix is rewriting it, not retraining six evaluators. Genesys publishes no threshold for acceptable variance, because the right spread depends on the form; that decision belongs to the client.

Genesys documents how to assign and index calibrations, but does not say that evaluators score blind to each other or to the expert. Sequencing is a discipline you impose, not a setting, and an expert's score is a reference point, not proof of correctness.

11

Automated scoring: Evaluation Assistance and AI scoring

Each question has an automation mode: No Automation, AI Scoring or Evaluation Assistance. It is a per-question choice with programme-wide consequences, and both automated modes support voice and digital interactions.

11.1 The three modes

Automation modes for evaluation questions (from Genesys)
ModeMechanismSuited toConstraints
No AutomationThe evaluator answersJudgement, and anything not in the transcriptNone
Evaluation AssistanceConditions built from speech and text analytics topicsDetectable events: was something said, did a topic fireUp to 10 topics per condition; the only automation for multi-select questions
AI ScoringGenerative AI over the form text and the transcriptBehaviours expressed in what was said or written50 per form; up to 5 answer options; 50 AI scoring requests per agent per day
  • Genesys's case for AI scoring is faster, more consistent evaluation, less human bias, a broader and more representative sample, and answer explanations for feedback.
  • Breadth of sample is usually the real business case. Genesys frames automated scoring in programs the same way: AI can score a far larger share of interactions than manual review, and quality teams spend more time on coaching.
  • Changing a question's mode is a form change and needs a new version, so scores before and after measure differently even with identical wording.

11.2 Evaluation Assistance

Genesys explains that Evaluation Assistance uses conditions made of topics, which represent business intent through collections of related phrases; its example is a Cancellation topic containing phrases such as close out my account. When speech and text analytics find a match, an answer is generated automatically. The feature arrived in January 2023.

  • A condition can include or exclude topics. For an include condition, any one topic found makes it true; for an exclude condition, all its topics must be absent.
  • When an answer has several conditions, all must be true for the answer to be entered.
  • A condition can be configured for only one answer per question, and a topic cannot be used in more than one condition.
  • Deleting a topic invalidates any condition that uses it.
  • No match means no prefill: the question is left for a human, or the default applies.

11.3 How AI scoring works

Since March 2025, AI scoring has used generative AI to answer evaluation questions. Genesys says the prompt is built automatically from the question group, the question, the help text and the answers. Wording, help text and answer labels are therefore all scoring logic, and an ambiguous label such as somewhat is ambiguous to the model exactly as it is to a person.

Genesys's best-practice article says AI scoring evaluates only the transcript and the form, and does not use metadata such as queue, routing path, timestamps, duration or system metrics, nor tone or inferred meaning. The Agent Auto-Evaluation article describes answers based on transcripts and metadata; we follow the more specific statement.

Genesys documents several further rules that affect whether, and how, an answer appears:

  • Genesys's FAQ states that AI scoring requires an interaction transcript, and the scoring article says it runs only when the transcript contains at least 300 characters.
  • A prompt filter means answers are provided only when the model's confidence is high; otherwise Genesys shows that there was not enough evidence.
  • The model's response is matched exactly against the answer options, so inconsistent accents or special characters across question, help text and answers can stop a valid answer matching.
  • With language auto-detection, reasoning text appears in the form's dialect, then the form's detected language, then the conversation language; the form's dialect can be set through the API.
  • Since May 2026, each automated question can focus on the evaluated agent or the full interaction, which matters for transfers and IVR-heavy interactions.

11.4 Permissions that AI scoring needs

Genesys lists the permissions required to use AI scoring in evaluations: Quality, Evaluation, View Sensitive Data; Speech and Text Analytics, Data, View; Recording, Recording, View; and Recording, RecordingSegment, View. View Sensitive Data also controls whether someone can see the reason for an AI answer.

Compare that with the default role lists on Genesys's roles page: neither default quality role lists View Sensitive Data, and the default Quality Evaluator role does not include Speech and Text Analytics, Data, View. An evaluator on unmodified defaults is therefore likely to see AI scoring behave as though it were absent. That is our reading of the two pages; confirm in your organisation.

12

Writing questions an AI can answer, and auto-evaluation

Genesys publishes a best-practice guide for AI-scored questions and a set of accuracy controls. Together with the Agent Auto-Evaluation form type, they decide whether automation helps a programme or quietly corrupts it.

12.1 Clear, objective, transcript-measurable

Genesys's aim is that every question is clear, objective and measurable from the transcript, and its sample form measures one observable behaviour per question. Its guidance: anchor on transcript evidence, make behaviours binary or specific, write complete sentences, define what counts and what does not, avoid questions combining several conditions, use agent and customer consistently, include short yes and no examples, and do not ask about tone or emotion.

The transcript test removes the most questions. Did the agent key the refund correctly, was the resolution the right one, did the agent check the account, was the customer satisfied: each depends on something the transcript may not contain.

12.2 Help text is the scoring rule

Genesys says help text should explain when each answer should be selected, not repeat the question. Its bad example asks whether the agent provided a complete resolution and tells the evaluator to choose accordingly; its good example defines Yes, No and N/A, including when N/A applies, such as a customer disconnecting before troubleshooting began. With 6,000 characters available, help text is where the decision rule, the boundary cases and the examples belong: the briefing you would give a new evaluator, written for a reader with no context.

Since June 2026, AI-scored questions can have up to five answer options, which Genesys says lets evaluations reflect degrees of compliance or empathy without extra questions. A graded scale often scores better than yes or no on anything involving degree.

12.3 Accuracy controls Genesys applies

AI scoring accuracy controls (from Genesys, checked 16 November 2026)
ControlHow Genesys describes it
Historical accuracyMeasured by how often humans modify AI answers, over the 25 most recent submissions per question; below 75%, the question is flagged low confidence and its AI answer hidden until accuracy recovers
History resetChanging and republishing a form resets the tracked history of AI-scored questions, and history is deleted after 30 days without activity
Agreement scoreShown per form and per question in the editor; Genesys suggests at least 20 test interactions, refined within a rolling 30-day window, and recommends aiming for 80% or more
Prompt filterAnswers only where model confidence is high; otherwise no answer, with context shown
Daily limit50 AI scoring requests per agent per day; Genesys's remedies are to spread evaluations, reduce retries or request a quota increase
  • Genesys's common errors also include a duplicate evaluation error, where an evaluation with the same evaluator and conversation already exists, and processing failures from vague or complex questions.
  • Because republishing resets history, a form that is revised every week never builds the history that lets AI answers show reliably.

12.4 Agent Auto-Evaluation forms

Genesys says Agent Auto-Evaluation forms are submitted automatically once every question has been answered by AI scoring, Evaluation Assistance or a default, without human review. Every question must use one of the two automated methods or the form cannot be published, every group needs a default answer, and every Evaluation Assistance question needs a condition. The evaluator is listed as Virtual Supervisor.

The fallback is where the risk sits. For AI-scored questions, Genesys says the default is used when the interaction is flagged low accuracy or the model's confidence is below 70%; for Evaluation Assistance, when no condition is met. With a default of the highest score, uncertainty is recorded as a pass, released and averaged like any other answer.

Since October 2026, analytics, interaction search and evaluation views include filters and labels that distinguish Virtual Supervisor evaluations from manual ones. Evaluations can be generated for one interaction through the conversation evaluations API, or recurringly through agent scoring rules in speech and text analytics programs, which arrived in February 2026.

12.5 When the model changes, scores move

Genesys announced on 12 January 2026 an update to the large language models behind Virtual Supervisor's AI scoring, with a planned effective date of 2 February 2026. It said results may change because the updated models interpret questions and evidence differently, that differences do not necessarily indicate errors, and recommended that customers review forms, validate results after the update, especially for high-impact or compliance questions, and adjust forms if needed.

So an AI scoring update can move a monthly average with no change in agent behaviour. Genesys used its announcements page to give about three weeks' notice of that change, which makes the announcements page part of a quality team's monitoring.

13

Reporting, coaching and retention

Reporting shows where the value appears, coaching is where it is realised, and retention decides whether a score can still be defended later.

13.1 Four levels of reporting

Where quality data appears (from Genesys)
LevelWhereWhat it shows
InteractionQuality Summary tabForm, form type, agent, evaluator, status, score, critical score, agent reviewed and commented, calibration flag, time remaining
AgentAgent Evaluations dashboard; Agents Evaluation Summary and Detail viewsForm, section, question and answer-level detail; Detail includes released and unreleased, policy and ad-hoc evaluations
EvaluatorQuality evaluator dashboard; Calibration Evaluations dashboardWorkload, average scores given, calibration activity and unfinished calibrations
APIEvaluation aggregate queries and jobs, plus a /me queryMetrics by queue, form or user, with analytics:evaluationAggregate:view
  • Since March 2026, score, critical score, evaluation type and evaluated agent appear directly in the interaction's quality summary table.
  • To see only Virtual Supervisor evaluations in the Detail view, Genesys says to filter by an Agent Auto-Evaluation form.
  • Genesys automatically adds Quality, Evaluation, searchAny to roles that hold Supervisor Agent Details view, Evaluation Edit or Evaluation Edit Score. searchAny is broad, so audit it on estates with division-scoped supervision.
  • Quality scores can be a gamification metric, which brings the six-day rescore rule and dispute updates into incentives; our gamification handbook covers the design.

13.2 Coaching: where quality pays or fails

Genesys coaching appointments bring together the interaction, recordings, guides and other material, with public and private notes for facilitator and agent, delivered to the agent's inbox. Appointments can be created from an interaction's detail view with Schedule Coaching, from workforce management, or from the Agent Development views, up to six weeks ahead. Scheduling offers optimal slots, which Genesys says use workforce management service goals, adherence and availability, as well as custom slots and off-schedule meetings. Coaching uses an activity code categorised as Meeting, Off Queue or Training, a capacity decision the workforce team should make.

Both default quality roles include Learning permissions, so a finding can become a learning module assignment in the same workflow. Supervisors can watch upcoming and overdue appointments, and agents see theirs in My Development.

13.3 Retention: indefinite scores, decaying evidence

Genesys keeps quality evaluation data indefinitely, until the organisation is terminated. Recordings follow the retention set in quality policies, subject to the maximum interaction data retention time: screen recordings have a one-year maximum, call and digital recordings no upper limit, and recordings older than the maximum interaction retention are deleted whether or not a policy covers them. Transcripts, topics and sentiment are deleted with the recording, and interactions older than 1.5 years have reduced detail and API functionality.

Set side by side, the risk is plain. An agent is scored on a compliance failure; the evaluation persists for years; the recording goes after 90 days. During an appeal two years later, the score exists and the evidence does not, so the agent's challenge cannot be tested. Our monitoring guide covers the wider lawful-basis questions.

  1. Align evaluation use with recording retention deliberately: extend retention for evaluated interactions, or limit the use of evaluations older than the recording window.
  2. Set a use limit, for example evaluations inform appraisal for 12 months, which also answers the data-minimisation question.
  3. Export what must be defensible, recordings and evaluation detail, before the recording expires.

14

Permissions, licensing, programme design and what goes wrong

The last part brings together the permission model, the licence dependencies, the order in which we design a programme, and the problems we see most often.

14.1 Default roles and the API

Genesys's default Quality Administrator holds forms, evaluations, calibrations, survey forms, recording retention and encryption keys, analytics views, Architect flow edit, coaching and all Learning assignment permissions. The default Quality Evaluator is narrower: it views forms; assigns, declines, edits scores, releases, searches and views evaluations; manages annotations; and plays, downloads and restores recordings. It cannot create forms or delete evaluations.

In the Platform API definition, read on 15 November 2026, there are 31 quality permissions and 19 coaching permissions. Edit Agent Sign-off maps to quality:evaluation:editAgentSignoff on the evaluation update operation, Edit AI scoring to quality:evaluationForm:aiScoringEdit on the form's AI scoring settings operation, and agent scoring rules use quality:scoringRule permissions. No operation in the definition requires a dispute or decline permission, although both appear on the roles page; how they are enforced is not documented.

  • The default Quality Evaluator role includes Recording, Encryption Key, Edit as well as View. Playing encrypted recordings is the legitimate reason for access; Edit is more than that needs.
  • Genesys states that a role's licence corresponds to the most expensive permission in it, so adding Architect or Learning permissions to a quality role can change what it costs to assign.
  • Permission conditions are finer than divisions: Genesys's example restricts a Quality Administrator's recording access to the Support queue.
  • The default Supervisor role includes two of the three permissions managers need to see evaluations; Quality, Evaluation, View has to be added.
  • Genesys notes that permission changes can take a few minutes to take effect.

14.2 Licensing dependencies

Genesys's summarised plan comparison, dated 3 November 2026, shows quality management and evaluations as an add-on on CX 1 and included on CX 2, CX 3 and CX 4, and shows Virtual Supervisor AI scoring as requiring experience units on all four tiers. Genesys's FAQ on AI scoring licences lists CX 1 with a WEM add-on and AI Experience units, CX 2 or CX 3 with AI Experience units, or CX 4. Older release notes carry narrower subscription lists, which is why entitlement should be read from the contract, not from a guide.

Genesys says Virtual Supervisor includes AI scoring, AI summary, insights and outline, and AI translate, metered under the AI Experience units model. So AI scoring is a consumption question even where quality management is included, and it must be sized. Evaluation Assistance depends on speech and text analytics and on transcription; quality depends on recordings existing. QVCCS does not sell licences; our AI Experience tokens guide explains the metering.

14.3 The order we design in

  1. Business decisions first: what is measured and by whom, fatal-question policy and its consequence, sampling design, division architecture, retention, dispute routing, notice for live monitoring, and the licence position on AI scoring.
  2. Forms: 25 questions or fewer, groups that match reporting, N/A never on critical questions, weights published with a worked example, disputes enabled, preview before publishing.
  3. A pilot with humans only: one team, one form, policy sampling and no automation, to establish evaluator agreement before any AI is involved.
  4. Automate the checklist: Evaluation Assistance on detectable events, then AI scoring on questions that pass the transcript test, each baselined against known interactions.
  5. Operate and prove it: weekly release backlog, coaching conversion, disputes and AI overrides; monthly calibration and control-set checks; quarterly coverage and fatal-question review.

14.4 What goes wrong

Common quality management problems, from our project experience
SymptomLikely causeFix
Two of three right, scored 60%Group weighting, correctly appliedPublish weights and a worked example (part 5)
Two agents score 90% on very different workN/A removed questions from the arithmeticRestrict N/A and require a comment (part 5)
Perfect critical score, key step unassessedCritical question marked N/ANo N/A on critical questions (part 5)
One omission gives a 0% evaluationFatal question answered NoReserve fatal for binary legal steps (part 4)
Weights revertedScaled scoring toggled off and onTreat the toggle as a scoring change (part 5)
Transfers evaluated against the wrong agentFirst-agent and last-agent rules differ by policyModel transfers; check new policies (part 7)
Evaluations done, agents see nothingNot releasedReport time-to-release weekly (part 9)
Agent cannot disputeAllow disputes off, permission missing, 45 days passed or maximum reachedCheck all four (part 9)
Average moved with no change in behaviourAI model update or rescoringControl set and annotated trends (part 12)
No AI prefills at allShort or missing transcript, or missing permissionsCheck transcript, then the four permissions (part 11)
AI answers vanished from a questionHistorical accuracy below 75%, or history reset by republishingRewrite the question and batch changes (part 12)
Managers cannot see evaluationsSupervisor role lacks Evaluation ViewAdd it, or a scoped role (part 14)
Division team cannot find the formForm Unassigned or in another divisionRecreate deliberately (part 6)
A score cannot be defendedEvaluation kept, recording expiredAlign retention and export (part 13)
Points awarded, then removedRescore saved but not submittedFinish or discard the same day (part 9)
Custom report shows strange statesDisputes enabled: In Review and Retracted appear in API resultsUpdate the report logic (part 9)

Questions

Common questions

How does Genesys Cloud calculate an evaluation score?

Genesys multiplies each question group's points by the group's weight and divides by each group's possible points times its weight. In Genesys's example, an agent with two of three questions right scores 60%, because the failed group carries more weight. Questions marked N/A show zero out of zero and leave the calculation, and a fatal question answered No sets the score to zero.

What is the difference between a critical and a fatal question in Genesys Cloud?

Both must be scored Yes or No. A critical question counts towards the overall score and a separate critical score. A fatal question is also a critical question, and if it is answered No the evaluation scores zero; in Genesys's worked example, both the overall and critical scores become 0%. We reserve fatal for one or two binary, legally required steps per form.

How long does an agent have to dispute a Genesys Cloud evaluation?

Genesys states that agents can open a dispute within 45 days, and only on completed and released evaluations. The form must have Allow disputes enabled, the agent needs the Dispute permission, and the form sets a maximum number of disputes per evaluation, up to 5. Separately, agents cannot review evaluations more than 90 days after the evaluation was created.

Why is Genesys AI scoring not prefilling answers?

Check the transcript first: Genesys says AI scoring requires one and runs only when it has at least 300 characters. Then check permissions: View Sensitive Data, Speech and Text Analytics Data View, Recording View and RecordingSegment View. Answers are also hidden for a question whose historical accuracy over its last 25 submissions falls below 75%, and republishing a form resets that history.

What happens when the confidence of an Agent Auto-Evaluation answer is low?

Genesys says the question's default answer is used when the interaction is flagged low accuracy or the model's confidence is below 70%, and the evaluation is still submitted automatically as Virtual Supervisor. If the default is the highest-scoring answer, uncertainty is recorded as a pass, so we point defaults at the least favourable or N/A answer, or require human review.

How long does Genesys Cloud keep quality evaluations?

Genesys keeps quality evaluation data indefinitely, until the organisation is terminated. Recordings follow the retention set in quality policies and the maximum interaction data retention time, and screen recordings have a one-year maximum. Transcripts are deleted with the recording, so a score can outlive the evidence behind it unless retention is aligned or the recording exported.

Sources

  1. AI-driven quality management scoring and assessment updates – Genesys Cloud Resource Centerhelp.genesys.cloud
  2. Automated agent scoring in programs for scalable evaluation – Genesys Cloud Resource Centerhelp.genesys.cloud
  3. About coaching – Genesys Cloud Resource Centerhelp.genesys.cloud
  4. About quality management – Genesys Cloud Resource Centerhelp.genesys.cloud
  5. About Virtual Supervisor – Genesys Cloud Resource Centerhelp.genesys.cloud
  6. Agent auto evaluation form – Genesys Cloud Resource Centerhelp.genesys.cloud
  7. Agent evaluations – Genesys Cloud Resource Centerhelp.genesys.cloud
  8. Agents Evaluation Detail view – Genesys Cloud Resource Centerhelp.genesys.cloud
  9. Agents Evaluation Summary view – Genesys Cloud Resource Centerhelp.genesys.cloud
  10. AI scoring best practices – Genesys Cloud Resource Centerhelp.genesys.cloud
  11. Allow evaluation disputes – Genesys Cloud Resource Centerhelp.genesys.cloud
  12. Automated scoring options in evaluation forms – Genesys Cloud Resource Centerhelp.genesys.cloud
  13. Calibration evaluations – Genesys Cloud Resource Centerhelp.genesys.cloud
  14. Calibration overview – Genesys Cloud Resource Centerhelp.genesys.cloud
  15. Conduct a calibration – Genesys Cloud Resource Centerhelp.genesys.cloud
  16. Copy an evaluation form – Genesys Cloud Resource Centerhelp.genesys.cloud
  17. Create evaluations by agents – Genesys Cloud Resource Centerhelp.genesys.cloud
  18. Create a new evaluation – Genesys Cloud Resource Centerhelp.genesys.cloud
  19. Create and publish an evaluation form – Genesys Cloud Resource Centerhelp.genesys.cloud
  20. Data retention periods and export functionality by feature – Genesys Cloud Resource Centerhelp.genesys.cloud
  21. Edit an evaluation form – Genesys Cloud Resource Centerhelp.genesys.cloud
  22. Evaluate in-progress interactions – Genesys Cloud Resource Centerhelp.genesys.cloud
  23. Evaluate an interaction – Genesys Cloud Resource Centerhelp.genesys.cloud
  24. Evaluation form weights – Genesys Cloud Resource Centerhelp.genesys.cloud
  25. Manage disputed evaluations – Genesys Cloud Resource Centerhelp.genesys.cloud
  26. My Evaluations view – Genesys Cloud Resource Centerhelp.genesys.cloud
  27. Agents Performance views overview – Genesys Cloud Resource Centerhelp.genesys.cloud
  28. Quality evaluator dashboard – Genesys Cloud Resource Centerhelp.genesys.cloud
  29. Quality management process – Genesys Cloud Resource Centerhelp.genesys.cloud
  30. Quality management roles and permissions – Genesys Cloud Resource Centerhelp.genesys.cloud
  31. Revise the score of an interaction – Genesys Cloud Resource Centerhelp.genesys.cloud
  32. Retention period for analytics data and recording – Genesys Cloud Resource Centerhelp.genesys.cloud
  33. Review a published evaluation form – Genesys Cloud Resource Centerhelp.genesys.cloud
  34. Schedule a coaching appointment – Genesys Cloud Resource Centerhelp.genesys.cloud
  35. Understand coaching appointments – Genesys Cloud Resource Centerhelp.genesys.cloud
  36. View a completed evaluation – Genesys Cloud Resource Centerhelp.genesys.cloud
  37. View an interaction's quality summary – Genesys Cloud Resource Centerhelp.genesys.cloud
  38. Review released evaluations – Genesys Cloud Resource Centerhelp.genesys.cloud
  39. Contact center managers home – Genesys Cloud Resource Centerhelp.genesys.cloud
  40. How can managers see their employee's quality evaluations? – Genesys Cloud Resource Centerhelp.genesys.cloud
  41. What does Virtual Supervisor include? – Genesys Cloud Resource Centerhelp.genesys.cloud
  42. What is the retention period for quality management evaluation data? – Genesys Cloud Resource Centerhelp.genesys.cloud
  43. Which evaluation policy should I use? – Genesys Cloud Resource Centerhelp.genesys.cloud
  44. calibration – Genesys Cloud Resource Centerhelp.genesys.cloud
  45. critical question – Genesys Cloud Resource Centerhelp.genesys.cloud
  46. fatal question – Genesys Cloud Resource Centerhelp.genesys.cloud
  47. quality admin – Genesys Cloud Resource Centerhelp.genesys.cloud
  48. August 30, 2023 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  49. December 18, 2019 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  50. February 2, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  51. Genesys Cloud FedRAMP region - August 18, 2025 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  52. Genesys Cloud FedRAMP region - January 31, 2024 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  53. January 25, 2023 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  54. January 26, 2022 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  55. July 10, 2019 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  56. June 1, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  57. June 22, 2022 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  58. March 2, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  59. March 24, 2025 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  60. May 25, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  61. October 25, 2023 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  62. October 5, 2026 release notes – Genesys Cloud Resource Centerhelp.genesys.cloud
  63. Employee Performance – Genesys Cloud Resource Centerhelp.genesys.cloud
  64. Quality Assurance and Compliance – Genesys Cloud Resource Centerhelp.genesys.cloud
  65. Does AI Scoring require a transcript? – Genesys Cloud Resource Centerhelp.genesys.cloud
  66. What licenses are required to use AI Scoring? – Genesys Cloud Resource Centerhelp.genesys.cloud
  67. Evaluation aggregates – Genesys Cloud Developer Centerdeveloper.genesys.cloud
  68. Genesys Cloud Platform API v2 definition (permission requirements per operation) – Genesysapi.mypurecloud.com
  69. Genesys Cloud CX summarised plan comparison – Genesyslibrary.genesys.com

Last reviewed

We work by introduction

Clients, partners and people introduced to us can reach the right specialist directly.

Who to contact